Fanli — China's fisheries large model, from 1.0 to 397 billion parameters in two years

East-Asia (China); developed at national level, opened globally

Content

Fanli (范蠡) — named for the ancient Chinese statesman associated with fish farming — is China’s fisheries large model, built at the National Digital Fisheries Innovation Centre at China Agricultural University under Li Daoliang (李道亮), professor in the School of Information and Electrical Engineering and dean of the International College.

The iteration record.

• Fanli 1.0 released at CAU on 15 June 2024, described as China’s first fisheries large model: multimodal fisheries data collection, cleaning, extraction and integration, aimed at fisheries workers and producers.
• Fanli 4.0 released 17–18 August 2026 at the 2026 International Smart Fisheries and Aquaculture Conference in Beijing (hosted by CAU with FAO support, 400+ participants). Reported specification: 397 billion parameters; coverage of eight core aquaculture dimensions — water quality, species, feed, health, operations, equipment, energy and economics; expert knowledge for 101 major farmed species; described as the industry’s most complete vertical-domain knowledge graph.

So: four named versions in 26 months, and a parameter count that grew from unstated to 397 billion. That trajectory is the unit’s most useful fact — the generational cadence of Chinese agricultural AI is fast, and none of it is published with evaluation.

The dataset is the more consequential release. Alongside 4.0, the centre published what it calls the world’s first public smart-fisheries dataset: 116,000 image frames and 698,000 task annotations, forming an image and visual-cognition dataset. Model and dataset were declared open to the world, with an explicit invitation for scholars and companies to use, evaluate, co-build and share — a framing that anticipates scrutiny rather than avoiding it.

The FAO layer. FAO Assistant Director-General and Fisheries and Aquaculture Division director Manuel Barange attended and spoke of continuing cooperation with the CAU centre around three cores — blue transformation, digital fisheries, sustainable fisheries — and five directions: technology empowerment, green production, scale promotion, talent co-building and global export. An FAO-supported Chinese fisheries model is a materially different export channel from the vendor hardware routes documented elsewhere in the China cycle.

What surrounds it. The conference ran three tracks — data, AI and decision-support systems; sensing, smart equipment and robotics; and integrated smart-production systems with whole-chain coordination. Chinese reporting also records the standards work moving underneath: a smart fishery farm (智慧渔场) construction industry standard, alongside the 《农业农村大模型》系列标准 — an agricultural-and-rural large-model standards series — which is the state’s attempt to make model inputs comparable in fisheries as elsewhere.

What this unit is doing in the taxonomy

The corpus’s first aquaculture unit anywhere, and its first Chinese generative-AI unit for a food-production sector. It anchors the generative AI × aquaculture cell that the activity matrix had left empty, and it is the Chinese counterpart to the corpus’s other domain-model work.

Distinct from:

• china-deep-sea-smart-aquaculture-platforms.md — the physical platform and farming-vessel layer; this is the model and data layer.
• china-livestock-ai-standards-and-research-infrastructure.md — CAAS’s HABLer and the Ministry’s livestock standards; same function (measurement infrastructure) in a different sector, different institution (CAU vs CAAS).
• The corpus’s agricultural-LLM units elsewhere — Fanli is a domain model with a released dataset, not a vendor assistant product.
• aquaculture-ai-type material in other jurisdictions — Canada’s aquaculture AI units are commercial farm-side systems; China’s is a national university centre releasing public data.

Why it matters for talks

• China released an aquaculture model and dataset to the world, under FAO auspices. That is the most open act in the corpus’s Chinese agri-AI record and it cuts against the closed-platform pattern that dominates the country’s commercial layer.
• 397 billion parameters for fisheries is a scale claim worth pausing on: it is far larger than most domain models in the corpus and is presented without evaluation. Parameter count is not capability; the unit records it as a cadence signal, not a performance claim.
• 116,000 image frames and 698,000 annotations is a concrete, checkable contribution — the kind of number that can be verified by downloading the dataset, unlike a deployment claim.
• The eight dimensions, 101 species structure shows what a genuinely domain-shaped model looks like — water quality, feed, health, equipment, energy and economics in one system, rather than a chatbot over agricultural text.
• FAO as a channel makes this the corpus’s clearest case of Chinese agri-AI moving through a multilateral technical institution rather than through bilateral deals.

Critical context

• No published evaluation exists for any Fanli version. The claim “industry’s most complete vertical-domain knowledge graph” is institutional self-description. maturity-verification: V0 applies to a university centre’s release as much as to a vendor’s.
• Fast iteration may indicate immaturity as much as progress — four versions in 26 months, with parameter counts used as the progress metric, is a pattern the corpus should treat skeptically in any sector.
• “World’s first” public fisheries dataset is a narrow claim — first public image-and-visual-cognition dataset for fisheries, not the first fisheries data release of any kind. Chinese “first” claims are typically specific and should be paraphrased as narrowly as they are made.
• Deployment to farms is the missing half. Nothing here shows a named farm, cooperative or region using Fanli in operations; the conference’s own tracks are technology-oriented. The unit reaches T3 on institutional pathway only.
• The dataset’s licence and access terms were not established — “open to the world” is a statement at a conference, not a licence. Check before relying on the data being reusable.
• 1.0’s release date (June 2024) and the FAO framing (2026) are two years apart — read the FAO involvement as a 2026 development, not as validation of the earlier versions.